Antsirabe (Madagascar)

Transcription

1 Antsirabe (Madagascar) JECAM/GEOGLAM Science Meeting Brussels, Belgium November, 2015 V.Lebourgeois, E.Vintrou, S.Dupuy, A.Bégué, J.Dusserre, M.Ameline, B.Bellon De La Cruz F.Ramahandry, C.Nativel

2 Site Description Location: 60*60km in the Highlands (Lat.: o S / Long.: o E) Topography: hills / plains / shallows ( m) Soils: Clayey Drainage class / irrigation: Moderate / irrigation canals Crop calendar: October - april Field size: 0.03 ha Climate and weather: Subtropical humid (cloudy!) Agricultural methods used: Manual Tillage / Hoeing / Fertilization with manure more or less mixed with ashes (few NPK inputs due to availability and cost) / Irrigation on terraces or basins, rainfed crops on the hills

3 Site Description Rainfed crops Irrigated crops

4 Project Objectives Crop identification and Crop Area Estimation Using existing missions providing decametric images (simulating Sentinel-2 data), VHSR imagery (PLEIADES) and other data (DEM, field database) With OBIA and data mining (Random Forest) Yield Prediction and Forecasting Analyzing time series to monitor crop phenology Estimating yield of rice crop (main crop of the Region, and key information for food security systems)

5 Collaboration Koumbia (Burkina Faso Cirad site) Similar methods used (object based image analysis and Random Forest) Comparisons in overall accuracies obtained for the classifications Comparisons in metrics used (which are the most informative variables?) Champion user of S2-Agri project SIGMA project

6 Earth Observation (EO) Data Received/Used 1rst growing season ( ): SPOT4 Take5 experiment SPOT5 (10m) from SEAS-OI receiving station DEIMOS (20m) from Deimos Imaging LANDSAT 8 (15m) from USGS PLEIADES (0.5m) from CNES HSR Time series VHSR coverage (pic of growing season) 2013 Growing season 2014 aug sept oct nov dec jan feb mar apr may jun 2 nd : 7 days in average between 2 acquisition of the HSR time series 2014 Growing season 2015 aug sept oct nov dec jan feb mar apr may jun 3 rd : 13 days in average between 2 acquisition of the HSR time series

7 In situ Data Land use and crop characterization GPS waypoints (with contour digitized on PLEIADES imagery) Attributes: Land use (crop or non crop), toposequence, irrigation, associated crops (if any), photo Rice yield GPS waypoints (with contour digitized on PLEIADES imagery) Attributes: Seeding and transplant dates, variety, fertilization (if any), toposequence, straw biomass, full grain weight, empty grain weight, photo

8 Land use and crop characterization In situ Data NON CROP

9 Rice yield In situ Data (88 fields) (36 fields)

10 Methods Crop identification and Crop Area Estimation Preprocessing Orthorectification and TOA reflectance calculation for decametric time series (simulating Sentinel-2 data) and VHSR imagery (PLEIADES) DEM processing (slope, watersystem extraction) Digitizing fields contour (based on PLEIADES imagery and GPS waypoints) Building a learning database 265 radiometric variables (mean and stdev spectral response in each band, indices, etc) from time series or PLEAIDES imagery 135 textural variables (mainly from PLEIADES imagery) 5 static variables (slope, field size, etc)

11 Methods Crop identification and Crop Area Estimation Random Forest Building a classifier (based on learning database) # Hierarchical method : 1. Crop/non crop classification 2. «Crop group» and «crop class» classification within the cropland only Analyzing the importance (informative degree) of the different variables Optimizing the classification (by reducing the number of variables used)

12 Accuracy Accuracy Accuracy Accuracy Results Crop identification and Crop Area Estimation CROP/NON-CROP Best OA = 93.8% 100% 80% 60% 40% 20% Overall Accuracy Crop Non crop CROP GROUP Best OA = 69.2% 0% 100% 80% 60% 40% 20% PLEIADES + STATIC METRICS TEMPORAL + STATIC METRICS ALL METRICS (PLEIADES + TEMPORAL + STATIC) Overall Accuracy Cereals Root/tuber crops Leguminous Other crops CROP CLASS Best OA = 58.9% 0% 100% 80% 60% 40% 20% 0% 100% 80% PLEIADES + STATIC METRICS PLEIADES + STATIC METRICS TEMPORAL + STATIC METRICS TEMPORAL + STATIC METRICS ALL METRICS (PLEIADES + TEMPORAL + STATIC) ALL METRICS (PLEIADES + TEMPORAL + STATIC) Overall Accuracy Bean Casava Corn Other crops Potatoe Rice Sweet potatoe Taro Irrigated rice Rainfed rice CROP SUB-CLASS 60% 40% 20% 0% PLEIADES + STATIC METRICS TEMPORAL + STATIC METRICS ALL METRICS (PLEIADES + TEMPORAL + STATIC)

13 Results Crop identification and Crop Area Estimation Optimization: how to map the whole area? (computer/volume limits) PLEIADES 0.5 m

14 OA Results Crop identification and Crop Area Estimation Optimization: reduce the volume of variables used 5/10 metrics for CROP/NON CROP (NDVI-SWIR-Panchromatic PLEIADES) 45 metrics for CROP CLASS (SWIR-NDVI-Panchromatic PLEIADES) 70% 60% 58.9% 60.2% 60.9% 59.3% 58.5% 58.7% 57.2% 53.6% 53.6% 52.7% 50% 48.2% 40% 30% 20% 10% 0% 405 var 50 var 45 var 40 var 35 var 30 var 25 var 20 var 15 var 10 var 5var Overall accuracies for crop class level according to number of variables used

15 Surveyed fields Methods Rice crop yield estimation

16 Methods Rice crop yield estimation Smoothing: - Stavitsky-Golay. Eliminate noise linked to clouds, mixed pixels, different sensors, atmospheric effects) Different tresholds: - Eliminate fields with incomplete temporal profil - Eliminate fields with smoothed profil to far from raw profil

17 Methods Rice crop yield estimation : Max NDVI of the growing season 2 à 12 : Integrals on different periods of the growing cycle

18 Valeur R² ou p-value Population Results Rice crop yield estimation full.grain.biom = f (integ.midmax) Données : TOA seuil.w = R² p-value population 80 The more the sorting is drastic, the more the correlations are good (but with less population!) Good linear correlations between some integrals and total biomass, straw biomass, acceptable for grain yield Currently not applicable to the whole production area but can give tendencies for food security systems % 25% 30% 35% 40% 50% 100% Quantiles RMSE (entre NDVI brut et lissé) 0

19 Research Plans for Next Growing Season Will you hold the course, or modify the approach? Multisource data analysis + Sentinel-2 data processing with a mixed image/signal processing and spatial modeling approach in order to be able to include expert knowledge on cropping systems/landscape organisation Do you anticipate using the same type/quantity of EO data next year? Y/N YES